Astria is a next-generation, fully local multimodal foundation model built on top of a Ministral-based language backbone and a custom vision encoder. This architecture significantly improves visual grounding, multilingual reasoning, and agentic reliability while remaining efficient enough for edge deployment.
🚀 Astria Update Highlights
Me7war’s latest Astria update pushes the limits of small-scale multimodal AI, combining efficiency, reasoning, and vision capabilities:
Key Features
Vision Mastery: Custom encoder enables deep image understanding and precise visual–text alignment.
A fully local, compact model redefining what edge-deployable multimodal AI can achieve.
📊 Visual Reasoning Performance
Astria Performance
Astria applies a custom evaluation using GPT-5 PRO as the judge.
92.53% — New SOTA
LLaVA baseline: 90.92%
A custom evaluation on 30 unseen images with 3 instruction types per image (conversation, description, complex reasoning) shows Astria outperforms GPT-5 in all categories.
Evaluation: Astria vs GPT-5
Astria Evaluation
A custom evaluation set of 30 unseen images was constructed. Each image includes three instruction types:
Conversational understanding
Detailed visual description
Complex multimodal reasoning
This yields 90 unique image–language tasks, evaluated on:
Astria
GPT-5
Scoring was performed by GPT-5 PRO, using a 1–10 scale per task.
Results
Astria outperforms GPT-5 across all instruction categories, validating the effectiveness of the custom vision encoder combined with the Ministral knowledge-enhanced language model.
Model Summary
Vision Encoder: Custom-built, with precise visual-text alignment
Language Backbone: Ministral-based, optimized for reasoning and factual accuracy
Training: End-to-end multimodal alignment with knowledge supervision
Output: Grounded, structured, and context-aware responses
Deployment: Fully local and edge-optimized, supporting up to 256k token context
License
Astria is released under the Astria License for personal and non-commercial use. Commercial use requires explicit permission from the creator.
